Access Type

Open Access Dissertation

Date of Award

January 2019

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

First Advisor

Cheng-Zhong Xu

Abstract

Cloud computing is becoming a fundamental facility of society today. Large-scale public or private cloud datacenters spreading millions of servers, as a warehouse-scale computer, are supporting most business of Fortune-500 companies and serving billions of users around the world. Unfortunately, modern industry-wide average datacenter utilization is as low as 6% to 12%. Low utilization not only negatively impacts operational and capital components of cost efficiency, but also becomes the scaling bottleneck due to the limits of electricity delivered by nearby utility. It is critical and challenge to improve multi-resource efficiency for global datacenters.

Additionally, with the great commercial success of diverse big data analytics services, enterprise datacenters are evolving to host heterogeneous computation workloads including online web services, batch processing, machine learning, streaming computing, interactive query and graph computation on shared clusters. Most of them are long-running workloads that leverage long-lived containers to execute tasks.

We concluded datacenter resource scheduling works over last 15 years. Most previous works are designed to maximize the cluster efficiency for short-lived tasks in batch processing system like Hadoop. They are not suitable for modern long-running workloads of Microservices, Spark, Flink, Pregel, Storm or Tensorflow like systems. It is urgent to develop new effective scheduling and resource allocation approaches to improve efficiency in large-scale enterprise datacenters.

In the dissertation, we are the first of works to define and identify the problems, challenges and scenarios of scheduling and resource management for diverse long-running workloads in modern datacenter. They rely on predictive scheduling techniques to perform reservation, auto-scaling, migration or rescheduling. It forces us to pursue and explore more intelligent scheduling techniques by adequate predictive knowledges. We innovatively specify what is intelligent scheduling, what abilities are necessary towards intelligent scheduling, how to leverage intelligent scheduling to transfer NP-hard online scheduling problems to resolvable offline scheduling issues.

We designed and implemented an intelligent cloud datacenter scheduler, which automatically performs resource-to-performance modeling, predictive optimal reservation estimation, QoS (interference)-aware predictive scheduling to maximize resource efficiency of multi-dimensions (CPU, Memory, Network, Disk I/O), and strictly guarantee service level agreements (SLA) for long-running workloads.

Finally, we introduced a large-scale co-location techniques of executing long-running and other workloads on the shared global datacenter infrastructure of Alibaba Group. It effectively improves cluster utilization from 10% to averagely 50%. It is far more complicated beyond scheduling that involves technique evolutions of IDC, network, physical datacenter topology, storage, server hardwares, operating systems and containerization. We demonstrate its effectiveness by analysis of newest Alibaba public cluster trace in 2017. We are the first of works to reveal the global view of scenarios, challenges and status in Alibaba large-scale global datacenters by data demonstration, including big promotion events like "Double 11".

Data-driven intelligent scheduling methodologies and effective infrastructure co-location techniques are critical and necessary to pursue maximized multi-resource efficiency in modern large-scale datacenter, especially for long-running workloads.

Download

Share

COinS